X-Vine Models for Multivariate Extremes

Anna Kiriliouk, Jeongjin Lee, Johan SegersJournal of the Royal Statistical Society Series B (Statistical Methodology) 2023

Abstract: Regular vine sequences permit the organisation of variables in a random vector along a sequence of trees. Regular vine models have become greatly popular in dependence modelling as a way to combine arbitrary bivariate copulas into higher-dimensional ones, offering flexibility, parsimony, and tractability. In this project, we use regular vine structures to decompose and construct the exponent measure density of a multivariate extreme value distribution, or, equivalently, the tail copula density. Although these densities pose theoretical challenges due to their infinite mass, their homogeneity property offers simplifications. The theory sheds new light on existing parametric families and facilitates the construction of new ones, called X-vines. Computations proceed via recursive formulas in terms of bivariate model components. We develop simulation algorithms for X-vine multivariate Pareto distributions as well as methods for parameter estimation and model selection on the basis of threshold exceedances. The methods are illustrated by Monte Carlo experiments and a case study on US flight delay data.

Introduction

Modeling extreme events in high dimensions—such as simultaneous financial market crashes or regional flooding—is notoriously difficult. Traditional Multivariate Extreme Value (MEV) models often suffer from the "curse of dimensionality," becoming either computationally intractable as the number of variables grows or too rigid to capture complex, asymmetric dependence patterns.

While "vine copulas" have revolutionized standard dependence modeling by breaking high-dimensional distributions into a sequence of linked bivariate trees, they could not be directly applied to extremes. This is because the "densities" used in extreme value theory (exponent measures) have infinite mass—they do not integrate to one—which breaks the standard probabilistic machinery used in central-limit modeling.

Methodology: The X-Vine Framework

The authors bridge this gap by developing the theory for X-vines (Extremal Vines). Their approach includes a new "Sklar’s Theorem" for tail densities, proving that tail copula densities can be decomposed into univariate margins and bivariate copula densities, provided they satisfy a specific homogeneity property.

The "secret sauce" of the paper is the Vine Telescoping Product Formula. This allows a $d$-dimensional tail copula density to be written as a product of $d-1$ bivariate tail copulas (the first tree) and $(d-1)(d-2)/2$ standard bivariate copulas (subsequent trees).

The first tree $T_1$ of a regular vine structure, representing the initial bivariate dependencies.

Instead of fitting a massive $d$-dimensional model all at once, the authors propose a tree-by-tree sequential estimation method using Maximum Likelihood. To keep models parsimonious, they use a modified Bayesian Information Criterion (mBIC) to decide when residual dependence is weak enough to stop adding trees, effectively truncating the vine.

The second tree $T_2$ of the vine, where nodes represent edges from the previous tree $T_1$.

Key Insights and Results

The authors show that many "classic" parametric models for extremes, such as the Hüsler–Reiss and Scaled Extremal Dirichlet models, are actually specific instances of X-vines. This provides a new theoretical lens to understand existing tools. Furthermore, X-vines allow for "mix-and-match" flexibility; researchers can use different bivariate models (e.g., a Gumbel copula for one pair and a Clayton for another) within the same high-dimensional structure.

The framework's scalability was tested with 50-dimensional models, demonstrating that X-vines can handle dimensions that would break traditional MEV approaches. In a case study of daily flight delays at 29 US airports, the X-vine model captured complex spatial dependencies better than standard graphical models, with the mBIC-optimal model truncating at level 7.

Limitations and Future Work

Despite its strengths, the sequential estimation method can lead to error propagation, where uncertainty accumulates at higher tree levels. Additionally, while the estimators work well in practice, the formal large-sample theory for this specific extremal context still needs to be fully developed. Future research may also explore extending this framework beyond max-stable distributions to conditional extremal models, which handle cases where variables do not necessarily become extreme simultaneously.

Summary updated

Abstract

Regular vine sequences permit the organisation of variables in a random vector along a sequence of trees. Regular vine models have become greatly popular in dependence modelling as a way to combine arbitrary bivariate copulas into higher-dimensional ones, offering flexibility, parsimony, and tractability. In this project, we use regular vine structures to decompose and construct the exponent measure density of a multivariate extreme value distribution, or, equivalently, the tail copula density. Although these densities pose theoretical challenges due to their infinite mass, their homogeneity property offers simplifications. The theory sheds new light on existing parametric families and facilitates the construction of new ones, called X-vines. Computations proceed via recursive formulas in terms of bivariate model components. We develop simulation algorithms for X-vine multivariate Pareto distributions as well as methods for parameter estimation and model selection on the basis of threshold exceedances. The methods are illustrated by Monte Carlo experiments and a case study on US flight delay data.
BibTeX
@article{kiriliouk2023x,
  author = {Anna Kiriliouk and Jeongjin Lee and Johan Segers},
  title = {{X-Vine Models for Multivariate Extremes}},
  journal = {Journal of the Royal Statistical Society Series B (Statistical Methodology)},
  year = {2023},
}
Figure